An Algorithm for Determining the Endpoints for Isolated Utterances

1. An Algorithm for Determining the Endpoints for Isolated Utterances L.R. Rabiner and M.R. Sambur The Bell System Technical Journal, Vol. 54, No. 2, Feb. 1975, pp. 297-315

2. Outline • Intro to problem • Solution • Algorithm • Summary

3. Motivation • Word recognition needs to detect word boundaries in speech • Recognizing silence can reduce: • Processing load • (Network not identified as savings source) • Relatively easy in sound proof room, with digitized tape

4. Visual Recognition • Easy • Note how quiet beginning is (tape) “Eight”

5. Slightly Tougher Visual Recognition • “sss” starts crossing the ‘zero’ line, so can still detect “Six”

6. Tough Visual Recognition • Eye picks ‘B’, but ‘A’ is real start • /f/ is a weak fricative “Four”

7. Tough Visual Recognition • Eye picks ‘A’, but ‘B’ is real endpoint • V becomes devoiced “Five”

8. Tough Visual Recognition • Difficult to say where final trailing off ends “Nine”

9. The Problem • Noisy computer room with background noise • Weak fricatives: /f, th, h/ • Weak plosive bursts: /p, t, k/ • Final nasals (ex: “nine”) • Voiced fricatives becoming devoiced (ex: “five”) • Trailing off of sounds (ex: “binary”, “three”) • With simple, efficient processing • Avoid hardware costs

10. The Solution • Two measurements: • Energy • Zero crossing rate • Show: simple, fast, accurate

11. Energy • Sum of magnitudes of 10 ms of sound, centered on interval: • E(n) = i=-50 to 50 |s(n + i)|

12. Zero (Level) Crossing Rate • Number of zero crossings per 10 ms • Normal number of cross-overs during silence • Increase in cross-overs during speech

13. The Algorithm: Startup • At initialization, record sound for 100ms • Assume ‘silence’ • Measure background noise • Compute average (IZC’) and stddev () of zero crossing rate • Choose zero-crossing threshold (IZCT) • Threshold for unvoiced speech • IZCT = min(25 / 10ms, IZC’ + 2 )

14. The Algorithm: Thresholds • Compute energy, E(n), for interval • Get max, IMX • Have silence, IMN I1 = 0.03 * (IMX – IMN) + IMN (3% of peak energy) I2 = 4 * IMN (4x silent energy) • Get energy thresholds (ITU and ITL) • ITL = MIN(I1, I2) • ITU = 5 * ITL

15. The Algorithm: Energy Computation • Search sample for energy greater than ITL • Save as start of speech, say s • Search for energy greater than ITU • s becomes start of speech • If energy falls below ITL, restart • Search for energy less than ITL • Save as end of speech • Results in conservative estimates • Endpoints may be outside

16. The Algorithm: Zero Crossing Computation • Search back 250 ms • Count number of intervals where rate exceeds IZCT • If 3+, set starting point, s, to first time • Else s remains the same • Do similar search after end

17. The Algorithm: Example (Word begins with strong fricative)

18. Algorithm: Examples • Caught trailing /f/ “Half”

19. Algorithm: Examples “Four” (Notice how different each “four” is)

20. Evaluation: Part 1 • 54-word vocabulary • Read by 2 males, 2 females • No gross errors (off by more than 50ms) • Some small errors • Losing weak fricatives • None affected recognition

21. Evaluation: Part 2 • 10 speakers • Count 0 to 9 • No errors at all

22. Evaluation: Part 3 • Your Project 1b…